Google · Available now

Gemma 4 by Google (2026): sizes from E2B to 31B, price per reply and context window

Gemma 4 is Google DeepMind’s family of open models, released on April 2, 2026 under Apache 2.0 in E2B, E4B, 26B and 31B sizes (a 12B followed on June 3) with context up to 256K tokens; the 31B is listed on OpenRouter at $0.20 per million input and $0.40 per million output tokens.

Available nowLast updated:

Use it now

Chat

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
Google
Type
Text
Released
2026-04-02
OpenRouter id
google/gemma-4-31b-it

Per million tokens

Price

Gemma 4 · checked 2026-11-21
-InputOutputSource
Via OpenRouter (google/gemma-4-31b-it)$0.20$0.40openrouter.ai

Checked

Facts, with sources

  • Google’s launch post, dated April 2, 2026, introduces Gemma 4 as "our most intelligent open models to date", built from the same research and technology as Gemini 3 and released under an Apache 2.0 license. It says the family launched in four sizes: Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture of Experts (MoE) and a 31B Dense, aimed at advanced reasoning and agentic workflows, with native function calling and structured JSON output.Source: blog.google · checked 2026-11-21
  • On June 3, 2026 Google added Gemma 4 12B. Its post says it sits between E4B and the 26B MoE, runs locally on 16GB of VRAM or unified memory, is the first mid-sized Gemma 4 with native audio input, and has no separate vision or audio encoders. Google says its benchmark performance is near the 26B model at less than half the total memory. These are the maker’s own claims, not tested by us.Source: blog.google · checked 2026-11-21
  • Google’s model card lists five sizes: E2B, E4B, 12B, 26B A4B and 31B. The small models (E2B, E4B) have a 128K-token context and the others 256K. All take text and image input and generate text; the card says E2B, E4B and 12B also take audio. It says all models have configurable thinking modes, native function calling and a native system role, and support more than 140 languages.Source: ai.google.dev · checked 2026-11-21
  • Google’s launch post says the weights are on Hugging Face, Kaggle and Ollama; the 31B and 26B run in Google AI Studio and the E2B and E4B in the Google AI Edge Gallery app; and Vertex AI, Cloud Run and GKE are available for production. It says the unquantized bfloat16 26B and 31B weights fit on a single 80GB NVIDIA H100 GPU, and the edge models run offline on phones, Raspberry Pi and NVIDIA Jetson Orin Nano.Source: blog.google · checked 2026-11-21
  • The model card header reads "License: Apache 2.0", and the launch post calls it a commercially permissive Apache 2.0 license, a change Google says followed developer feedback. The Hugging Face repositories google/gemma-4-31B-it and google/gemma-4-E2B-it show license apache-2.0 and are not gated (checked November 21, 2026).Source: ai.google.dev · checked 2026-11-21
  • Google publishes its own quantization-aware 4-bit (q4_0) GGUF files on Hugging Face for each size, in repositories named google/gemma-4-<size>-it-qat-q4_0-gguf (Apache 2.0, not gated): E2B 3.12 GiB, E4B 4.8 GiB, 12B 6.5 GiB, 26B A4B 13.45 GiB and 31B 16.44 GiB for the model file, plus a separate multimodal projector file of 0.16 to 1.12 GiB (sizes read November 21, 2026). We have not run any of them. Other uploaders have posted their own GGUF files, which we did not check.Source: huggingface.co · checked 2026-11-21
  • Google’s Gemini API pricing page lists Gemma 4 input, output and context caching as "Free of charge" on the free tier and "Not available" on the paid tier, and says content on the free tier is used to improve Google’s products (paid tier: no). We did not read the free-tier rate limits. Because Google lists no paid Gemma 4 price, this page has no maker price in dollars per million tokens; paid hosting comes from other providers.Source: ai.google.dev · checked 2026-11-21
  • OpenRouter lists google/gemma-4-31b-it (text, image and video input, text output, 262,144-token context) at $0.20 per million input and $0.40 per million output tokens, and google/gemma-4-26b-a4b-it at $0.0675 / $0.225. Individual providers charge differently: for the 31B, input runs from $0.10 to $0.75 per million across more than a dozen providers. E2B, E4B and 12B are not listed on OpenRouter (checked November 21, 2026). OpenRouter also lists ":free" variants of both models; we did not test their limits.Source: openrouter.ai · checked 2026-11-21 · second-hand
  • Google’s launch post says the 31B ranked #3 among open models and the 26B #6 on the Arena AI text leaderboard, in a chart noted "as of 4/1". That is the maker’s launch-day claim, not tested by us, and the ranking may have changed since.Source: blog.google · checked 2026-11-21

Sizes

Gemma 4 sizes side by side

  • The card gives E2B 2.3B effective parameters (5.1B with embeddings), E4B 4.5B effective (8B with embeddings), 12B 11.95B, 26B A4B 25.2B total with 3.8B active (8 of 128 experts plus one shared) and 31B 30.7B dense. The "E" means "effective": the small models use per-layer embeddings, so their effective count is much smaller than their total.Source: ai.google.dev · checked 2026-11-21

Side by side

Gemma 4 E2B, E4B, 12B, 26B A4B and 31B compared

People who search for Gemma 4 E2B or "gemma 4" usually want to know which size is which. E2B and E4B are the small ones for phones and laptops; 12B is the mid-size laptop model added in June; 26B A4B and 31B are the two larger ones. Only the 26B A4B and the 31B are listed on OpenRouter, so the chat on this page runs the 31B. For other open-weight models that have their own pages, see Qwen3.8 and Hy4 Flash.

Our table, from Google’s launch posts and model card, Hugging Face file listings and OpenRouter’s model list as read on November 21, 2026; credits are our arithmetic · checked 2026-11-21
WhatE2BE4B12B26B A4B31B
Released by GoogleApril 2, 2026April 2, 2026June 3, 2026April 2, 2026April 2, 2026
Parameters (model card)2.3B effective (5.1B with embeddings)4.5B effective (8B with embeddings)11.95B25.2B total, 3.8B active (MoE)30.7B dense
Context window (model card)128K tokens128K tokens256K tokens256K tokens256K tokens
Input (model card)Text, image, audioText, image, audioText, image, audioText, imageText, image
Google’s 4-bit GGUF model file (our check)3.12 GiB4.8 GiB6.5 GiB13.45 GiB16.44 GiB
OpenRouter price per million tokens (input / output)Not listedNot listedNot listed$0.0675 / $0.225$0.20 / $0.40
Chat on this site (our arithmetic)NoNoNoNoYes, on this page; a typical reply about 1 credit

Source: blog.googleSource: blog.googleSource: ai.google.devSource: huggingface.coSource: huggingface.coSource: huggingface.coSource: huggingface.coSource: huggingface.coSource: openrouter.ai

Nearest available

Models you can use now instead

Qwen3.8

Alibaba

$4 / $12 (per million tokens)

  • Alibaba announced Qwen3.8-Max on August 3, 2026, available through the Alibaba Cloud Model Studio API; it described it as the most powerful model in its Qwen series to date.Source: alibabagroup.com · checked 2026-11-14
  • Qwen3.8-Max is a sparse mixture-of-experts model with 2.4 trillion total parameters, about 95 billion active per inference, and a context window of up to 1 million tokens.Source: alibabagroup.com · checked 2026-11-14
  • Open weights: Qwen3.8-2.4T-A95B went up on Hugging Face and ModelScope on August 12, 2026, and Qwen3.8-27B on August 14, 2026.Source: github.com · checked 2026-11-14

Hy4 Flash

Tencent

$0.04 / $0.14 (per million tokens)

  • Tencent’s Hy4 flash announcement, dated November 16, 2026, presents it as the newest member of the Hy model family. Tencent recommends it for most everyday tasks (deep research, document processing, PowerPoint generation, simple front-end work, computer use) and points complex tasks such as long-horizon software engineering to Hy4. Tencent Cloud’s API guide calls it "a leaner, more efficient everyday workhorse".Source: tencent.com · checked 2026-11-17
  • The Tencent announcement is dated November 16, 2026, and Tencent Cloud’s model, pricing and API pages for hy4-flash were last updated the same day. The announcement says only that Hy4 flash is available now; it gives no separate release date or version name.Source: tencent.com · checked 2026-11-17
  • Tencent says Hy4 flash has 224B total parameters (including 64B Engram parameters), 12B active parameters, a context of up to 1M tokens and multimodal input; the announcement does not state a maximum output or the output types. It says the KV cache is 1/6 the size of Hy4’s and that decoding exceeds 150 tokens per second. These are the maker’s own figures, not tested by us.Source: tencent.com · checked 2026-11-17

FAQ

Questions

What is Gemma 4?

Gemma 4 is Google DeepMind’s family of open models. Google introduced it on April 2, 2026 in four sizes (E2B, E4B, a 26B mixture-of-experts called 26B A4B, and a 31B dense model) and added a 12B on June 3, 2026. Google calls it "our most intelligent open models to date", built for reasoning and agent use; all sizes take text and image input and write text, and E2B, E4B and 12B also take audio.

How much does Gemma 4 cost?

The weights can be downloaded without a fee under Apache 2.0; running them costs whatever your hardware or host charges. Google’s Gemini API pricing page lists Gemma 4 as free of charge on the free tier and not available on the paid tier, and says free-tier content is used to improve Google’s products. OpenRouter lists google/gemma-4-31b-it at $0.20 per million input and $0.40 per million output tokens, and the 26B A4B at $0.0675 / $0.225 (checked November 21, 2026); other providers charge differently. On this site a typical reply to this model costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit.

How do I use Gemma 4?

Google says the weights are on Hugging Face, Kaggle and Ollama, that the 31B and 26B A4B run in Google AI Studio and the E2B and E4B in the Google AI Edge Gallery app, and it lists Vertex AI, Cloud Run and GKE for production. On OpenRouter the models are google/gemma-4-31b-it and google/gemma-4-26b-a4b-it. You can also chat with the 31B in the chat box on this page, on credits.

What is Gemma 4 E2B (gemma 4 e2b)?

E2B is the smallest Gemma 4. The "E" stands for "effective": Google’s model card gives 2.3B effective parameters (5.1B with embeddings), a 128K-token context and text, image and audio input. Google says E2B and E4B run offline on phones, Raspberry Pi and NVIDIA Jetson Orin Nano. Google’s own 4-bit GGUF file for E2B is 3.12 GiB, plus a 0.92 GiB multimodal projector file; we have not run it. E2B is not on OpenRouter, so it cannot be chatted with here: the chat on this page runs the 31B.

What is the context window of Gemma 4?

Google’s model card gives 128K tokens for E2B and E4B and 256K tokens for 12B, 26B A4B and 31B. OpenRouter shows 262,144 tokens for the 26B A4B and 31B listings; the maximum output depends on the provider.

Is Gemma 4 open source, and can I run it locally?

Google released Gemma 4 under the Apache 2.0 license (model card and launch post), and the Hugging Face repositories we checked on November 21, 2026 show apache-2.0 and are not gated. Google says the unquantized 26B and 31B weights fit on a single 80GB NVIDIA H100, that quantized versions run on consumer GPUs and that the 12B runs on a laptop with 16GB of VRAM or unified memory. Google’s own 4-bit GGUF files are 3.12 GiB (E2B), 4.8 GiB (E4B), 6.5 GiB (12B), 13.45 GiB (26B A4B) and 16.44 GiB (31B); we have not run them. The chat on this page uses a hosted model, not a download.

How do the Gemma 4 sizes differ?

By Google’s model card, E2B and E4B are small models for phones and laptops (128K context, audio input); the 12B is a model without separate image or audio encoders (11.95B parameters, 256K context, audio input); the 26B A4B is a mixture-of-experts model with 25.2B total and 3.8B active parameters; the 31B is a 30.7B dense model with text and image input. Google says the 12B performs near the 26B on standard benchmarks at less than half the memory, and that the 31B ranked #3 among open models on the Arena AI text leaderboard at launch. Both are the maker’s own statements, not tested by us.

Is this the official Google website?

No. This is an unofficial page, not affiliated with or endorsed by Google. Facts link to Google’s own pages where possible.